Data and code for article "Nature reserve customized method of photo and video camera traps materials processing using two-stage neural network approach"

DESCRIPTION 📓

"data" folder directory contains the datasets for classification and detection. 

1. The detection dataset has YOLOv5 format and contains three classes [tigers, leopards, empty]. The class empty is about 10% of the total data. The leopard and tiger classes contain 3500 images each. The entire amount of data for the detection task is 7600 images.
2. The classification dataset contains two classes [tigers, leopards]. Images for classification are cropped images from the detection task using bounding boxes. Each class has 3500 images

The "weights" folder contains pretrained models for classification and detection tasks. 

• The detector weights were pre-trained on 231k images from camera traps located throughout Russia.
• The classifier weights were pre-trained on 416k images that were cropped with bounding boxes from photographs for the detection task. Some of the images for the classification task were taken from the Internet. The classifiers were trained for 29 classes.
• You can also find folder tigers_vs_leopards in both the detection and classification directory, where there are weights that have been trained on a part of the camera trap images available at the link below.

Classification weights

1. EfficientNetv2-M
2. ResNeSt-101e (🚀 RECOMMENDED)
3. ResNet-101d
4. ReXnet-100
5. SeResNet-152d

Detection weights

1. YOLOR-W6-1280
2. YOLOX-X-640
3. YOLOv5-X-640
4. YOLOv5-X-1280
5. YOLOv5-M6-1280
6. YOLOv5-L6-1280 (🚀 RECOMMENDED)

Read README.md file for more details